Trust · Cost attribution OTel · DEC-V11-80 · Q4-2026 GA

Cost attribution 4-layer + 6-dim · OTel GenAI semconv stable

4-layer cost attribution (prompt + tool + memory + response) · 6-dim (tenant · agent · module · model · provider · channel) · FOCUS 1.3 exporter · investor diligence-grade narrative canonical.

Status · preparation phase · Q4-2026 GA · DEC-V11-67 OTel migration
Section 1

4-layer cost attribution canonical

ROSS atribuye costo LLM por 4 capas independientes · vs single opaque cost competitors (Vapi · Bland · Retell · Artisan · 11x · todos billing per-minute o per-token sin granularidad layer). Diligence-grade transparency.

Layer 1 · Prompt
genai.prompt.tokens
Input tokens billable · system prompt + user message + RAG context + few-shot examples. Atribuido per agent + tenant + module.
Layer 2 · Tool
genai.tool.invocations
Function calling overhead · tool_use blocks · MCP server invocations · API external calls. Atribuido per tool_name + agent.
Layer 3 · Memory
genai.memory.retrievals
Vector retrieval cost · embedding generation · cache hits vs misses · Letta + mem0 + Graphiti queries. Atribuido per memory_layer.
Layer 4 · Response
genai.response.tokens
Output tokens billable · streaming chunks · thinking tokens (Opus 4.7). Atribuido per agent + model + tenant.
Section 2

6-dim attribution matrix

Cada token + cada tool call + cada memory retrieval atribuido en 6 dimensiones ortogonales · queryable OTel GenAI semconv stable standard (DEC-V11-67 migración Langfuse → OTel-native).

  • tenant_id · multi-tenant per-customer attribution
  • agent_slug · per-agent cofounder spend (Sara · Maya · Will · Patxi · Sales Cdr · Will Fund · Atlas · Irati · FLOW)
  • module_id · per-module cost (M01-M12 + M00 Brain)
  • model · per-LLM (Sonnet 4.6 · Opus 4.7 · Haiku 4.5 · GPT · Gemini)
  • provider · Anthropic · OpenAI · Google · self-hosted
  • channel · voice · email · LinkedIn · WhatsApp · cockpit

Permite query como "show me Sara voice spend Q3 vs Q4 break-down por module + tenant tier" o "drift detection per-layer past 30 days". Investor diligence canonical.

Section 3

FOCUS 1.3 exporter · industry standard

FinOps Open Cost & Usage Specification (FOCUS) 1.3 · cross-cloud cost data normalized standard FinOps Foundation. ROSS exporta cost data GET /api/finops/focus/v1.3 · JSON + NDJSON canonical.

FOCUS 1.3 row
BillingPeriodStart · BillingPeriodEnd
ServiceName · ServiceCategory
ResourceId · ResourceType
ChargeCategory · ChargeDescription
BilledCost · EffectiveCost · ListCost
+ROSS-tags · tenant · agent · module · layer · dim

Patxi FinOps cockpit consume FOCUS internally · investor diligence export same endpoint · auditor Big-4 friendly format · cross-cloud (Vercel · Supabase · Hetzner · Cloudflare · Anthropic · OpenAI · ElevenLabs) unified.

Section 4

genai_4layer_drift · canonical metric

Showcase chart placeholder
genai_4layer_drift
Per-layer cost drift detection · 30d rolling baseline · alert threshold +20% sustained 7d

Drift detection per-layer canonical · prompt drift (RAG bloat) · tool drift (new MCP integration cost) · memory drift (cache miss rate spike) · response drift (model upgrade Opus thinking tokens). Patxi alerts auto-fired · investor narrative · NO surprise burn.

Investor diligence · FinOps walkthrough

Pre-Series A investor diligence · FOCUS 1.3 export walkthrough + unit economics break-down per-layer + per-agent · 30 min.

Book ROSS Hour · FinOps walkthrough
AI Act art 50 disclosure · contenido generado con asistencia IA (Claude · Anthropic) · revisión humana firmada Will CTO + Patxi Compliance · DEC-V11-80.